It is a Rust CLI that hooks into Claude Code's lifecycle. Before large edits or before Claude finishes, another LLM evaluates the work against one question: is this the simplest thing that actually solves the stated problem?

When something's off, Claude is blocked and shown feedback like:

"This may be a local maximum. What alternatives were considered?"

"Is this necessary right now? This seems to solve [hypothetical problem] rather than [actual need]."

Claude then tries to self-corrects, often proposing something simpler or asking clarifying questions it should have asked upfront.

The philosophy: Every line of code is a liability. Every abstraction is a loan. If you can't explain it simply, it's too complex. If it feels clever, be suspicious.

Tradeoffs: Yes, this adds LLM calls. In my experience the tighter results are worth it, but YMMV.

Install: cargo install superego # or brew install cloud-atlas-ai/superego/superego

  sg init && claude

Source-available (free for personal/research use). GitHub: https://github.com/cloud-atlas-ai/superego